[Paper Review] Task-Adaptive Negative Class Envision for Few-Shot Open-Set Recognition

Shiyuan Huang, Jiawei Ma+2 more|arXiv (Cornell University)|Dec 23, 2020

Domain Adaptation and Few-Shot Learning41 references4 citations

TL;DR

This paper proposes task-adaptive negative class envision for few-shot open-set recognition (FSOR), dynamically generating negative prototypes from few-shot examples to learn adaptive rejection boundaries, eliminating manual threshold tuning. The method improves performance on both FSOR and generalized FSOR benchmarks by integrating threshold adaptation into the learning process through class correlation-aware negative generation.

ABSTRACT

We study the problem of few-shot open-set recognition (FSOR), which learns a recognition system capable of both fast adaptation to new classes with limited labeled examples and rejection of unknown negative samples. Traditional large-scale open-set methods have been shown ineffective for FSOR problem due to data limitation. Current FSOR methods typically calibrate few-shot closed-set classifiers to be sensitive to negative samples so that they can be rejected via thresholding. However, threshold tuning is a challenging process as different FSOR tasks may require different rejection powers. In this paper, we instead propose task-adaptive negative class envision for FSOR to integrate threshold tuning into the learning process. Specifically, we augment the few-shot closed-set classifier with additional negative prototypes generated from few-shot examples. By incorporating few-shot class correlations in the negative generation process, we are able to learn dynamic rejection boundaries for FSOR tasks. Besides, we extend our method to generalized few-shot open-set recognition (GFSOR), which requires classification on both many-shot and few-shot classes as well as rejection of negative samples. Extensive experiments on public benchmarks validate our methods on both problems.

Motivation & Objective

• Address the challenge of threshold tuning in few-shot open-set recognition (FSOR), where manual calibration is unreliable across diverse tasks.
• Overcome the limitations of traditional large-scale open-set methods in low-data regimes typical of FSOR.
• Improve rejection performance by learning dynamic, task-specific rejection boundaries that adapt to the few-shot context.
• Extend the framework to generalized few-shot open-set recognition (GFSOR), where both few-shot and many-shot classes are present alongside unknowns.
• Integrate negative prototype generation into the classifier training process to endow the model with inherent sensitivity to unknown samples.

Proposed method

• Augment the few-shot closed-set classifier with additional negative prototypes derived from the support set of each few-shot task.
• Generate negative prototypes using class correlations among few-shot examples to ensure semantic relevance and diversity.
• Learn dynamic rejection boundaries by optimizing the classifier to distinguish between known classes and the dynamically generated negative prototypes.
• Formulate the rejection boundary as a function of the learned negative prototypes, enabling task-specific adaptation without external threshold tuning.
• Extend the framework to GFSOR by jointly modeling known few-shot and many-shot classes while maintaining robust unknown sample rejection.
• Train the model end-to-end using a unified objective that combines classification and open-set discrimination via the generated negative prototypes.

Experimental results

Research questions

• RQ1 Can negative prototype generation from few-shot examples enable dynamic, task-adaptive rejection boundaries without manual threshold tuning?
• RQ2 How does incorporating class correlations in negative prototype generation improve generalization and rejection performance in few-shot open-set recognition?
• RQ3 To what extent does the proposed method generalize to the more complex generalized few-shot open-set recognition (GFSOR) setting?
• RQ4 How does the performance of the proposed method compare to existing FSOR methods that rely on fixed thresholding strategies?
• RQ5 Does the integration of negative class envision into the learning process lead to more robust and consistent open-set recognition across diverse few-shot tasks?

Key findings

• The proposed method achieves state-of-the-art performance on public few-shot open-set recognition benchmarks by learning task-adaptive rejection boundaries.
• By generating negative prototypes using class correlations, the model achieves improved generalization and robustness to unknown samples across diverse few-shot tasks.
• The method eliminates the need for manual threshold tuning, which is a major limitation in existing FSOR approaches.
• In the generalized few-shot open-set recognition (GFSOR) setting, the method maintains strong performance on both few-shot and many-shot classes while effectively rejecting unknown samples.
• Extensive ablation studies confirm that negative prototype generation based on class correlations significantly improves rejection accuracy compared to random or non-correlated negative sampling.